Large Language Model Agent in Financial Trading: A Survey (2408.06361v1)

Abstract: Trading is a highly competitive task that requires a combination of strategy, knowledge, and psychological fortitude. With the recent success of LLMs(LLMs), it is appealing to apply the emerging intelligence of LLM agents in this competitive arena and understanding if they can outperform professional traders. In this survey, we provide a comprehensive review of the current research on using LLMs as agents in financial trading. We summarize the common architecture used in the agent, the data inputs, and the performance of LLM trading agents in backtesting as well as the challenges presented in these research. This survey aims to provide insights into the current state of LLM-based financial trading agents and outline future research directions in this field.

• The paper reviews 27 studies to assess the performance and methodologies of LLM-based trading agents.
• It categorizes agents into 'LLM as a Trader' and 'LLM as an Alpha Miner', detailing the use of numerical, textual, visual, and simulated data.
• The survey evaluates performance using metrics like cumulative returns and Sharpe ratios, while noting research gaps such as limited testing periods and closed-source models.

LLM Agent in Financial Trading: A Survey

The paper "LLM Agent in Financial Trading: A Survey" provides an analytical review of the current landscape of utilizing LLMs as agents in the financial trading sector. In this comprehensive survey, 27 papers have been meticulously reviewed, signaling its aim to collate existing methodologies, evaluate their performance, and highlight future research trajectories within this domain.

Architecture of LLM-based Trading Agents

The survey categorizes LLM-based trading agents into two primary architectures: LLM as a Trader and LLM as an Alpha Miner.

• LLM as a Trader: This class of agents is designed to convert vast quantities of external data, such as news articles and financial reports, into actionable trading decisions like "Buy", "Hold", and "Sell". The paper identifies sub-types including news-driven, reflection-driven, debate-driven, and reinforcement learning-driven agents. For example, news-driven architectures primarily leverage sentiment analysis to forecast stock movements, while reflection-driven methodologies enhance decision-making by inferring reflections from accumulated memory.
• LLM as an Alpha Miner: These agents focus on generating alpha factors, which are numerical representations used in trading strategies, discerned by the LLMs. This is typically achieved through a system of inner-loop and outer-loop mechanisms, as demonstrated by architectures like QuantAgent and AlphaGPT, which refine the trading logic through iterative feedback from the market.

Data Utilization

The data utilized by these LLM agents are broadly categorized into numerical, textual, visual, and simulated data, each contributing uniquely to building comprehensive trading models:

• Numerical and Textual Data: Predominantly used by all LLM-powered agents, these include stock prices and alternative data such as news articles and financial reports. Textual data is processed to extract sentiment or insights that aid decision-making.
• Visual Data: Although less common, integrating visual data like trading charts shows promise in enhancing model comprehension of market dynamics.
• Simulated Data: Employed to replicate market scenarios for understanding agent behavior and testing performance without real-world risk.

Performance Evaluation

The surveyed literature highlights the agents' performance through rigorous backtesting, utilizing financial metrics such as cumulative returns, Sharpe ratios, and maximum drawdowns. Despite superior backtesting results, it is noted that the reliability of these findings varies given the often limited temporal scope of data used for evaluation.

Limitations and Future Directions

The paper acknowledges several limitations in the current research, such as the prevalence of closed-source models, challenges in integration with existing financial systems, and under-explored areas like real-time social media data impacts and fine-tuning efficacy. The reliance on backtests confined predominantly to the US and Chinese markets, alongside short testing windows, suggests an urgent need for more robust validations. Addressing these gaps could significantly optimize LLM applications in financial trading and extend their utility across varied financial instruments and global markets.

Conclusion

This survey paper consolidates knowledge on LLM-based trading agents, offering crucial insights into their current capabilities and limitations. As the integration of AI into financial trading systems continues to be explored, the findings of this survey offer a critical foundation, pointing towards potential advancements that could shape future developments in AI-enhanced financial trading. The continuous evolution of LLMs and their application in asset management presents a fertile ground for future research, which could redefine investment strategies and automation in financial markets.